2026-10-09 校核：Ava 原 GitHub 地址本次返回 404；这不能区分迁移、转私有或删除。机构、许可、架构代码、性能和引述均待原始材料复核，不能据星数断定采用规模或研发机构。本文配置建议尚未执行。 1

Ava — Many-Agent Civilization Simulation Framework

Source: GitHub — TianhangZhuzth/Fundamental-Ava (400⭐, Python, Apache 2.0) 机构归属：原笔记归属 Fundamental Research Labs，但尚无可复查的官方证据 Core premise: Runs large populations of autonomous agents — each with its own memory, belief system, and social model — inside a shared environment and detects emergent civilization-level properties.

项目定位

Ava is a research infrastructure project for studying emergent civilization-level behavior from large populations of cognitive agents. Unlike most multi-agent demos that top out at a handful of agents, Ava is designed for scale (100-5000+ agents) with three architectural bets:

1. Concurrency is structural: Every tick runs through asyncio.TaskGroup with a bounded asyncio.Semaphore, so a population of agents steps in parallel without one slow agent blocking the tick.
2. Memory is tiered: Episodic, semantic, and procedural memory are separate stores with different retrieval and decay semantics.
3. Emergence is measured statistically: Change-point detection (Mann-Whitney U) against population metrics — 原笔记列出 p-values，但没有原始数据、实验设计和多重检验信息，尚不能复核结论.

技术架构

Civilization
├── Culture (norms) / Governance (laws/votes) / EmergenceDetector (analysis)
├── ExecutionEngine (TaskGroup + Semaphore)
│   └── Agent 1..N (Memory + Belief + Social)
└── MessageBus (pub/sub)

关键设计决策

Tiered Memory Architecture

Each agent’s MemoryStore separates three types:

• Episodic: What happened (event stream with timestamps, recency-weighted retrieval)
• Semantic: What it means (distilled facts from reflection over important episodes)
• Procedural: How to do it (reinforcement-based skill learning, independent of episodes)

Quote from README:

“Procedural memory reinforces successful action sequences independently, so an agent that keeps succeeding at foraging in forest biomes converges on that skill without anyone hand-tuning a policy.”

Governance and BFT Consensus

Laws are proposed, voted on, and ratified by agents through the same GovernanceSystem, with quorum and majority thresholds scaling with population size. For adversarial scenarios, a three-phase PBFT protocol (PROPOSE → PREPARE → COMMIT) commits once 2f + 1 of n = 3f + 1 participants agree.

Statistical Emergence Detection

Rather than eyeballing “it looks like agents specialized”, the analysis layer runs Mann-Whitney U tests over a rolling window:

“role_specialization shifted from 0.140 to 0.710 (p=0.0021)“

Performance

对 Hermes/Manas 的意义

1. Tiered Memory Architecture: Ava’s episodic/semantic/procedural memory separation suggests a conceptual comparison with gbrain (pages/chunks/takes), rather than validating that implementation. The concept of procedural memory (reinforced skill learning) is an interesting extension — could inform how Hermes skills are prioritized based on successful usage patterns.

2. Emergence Detection: The statistical emergence detection (change-point + Mann-Whitney U) could be applied to detect coordination patterns across Alaya/Anchor/Mano — e.g., detecting when cross-agent communication patterns shift.

3. BFT Consensus for Agent Coordination: The PBFT-style protocol for agent consensus has direct relevance to multi-agent decision-making where agents may disagree or fail. The 2f+1 of n=3f+1 quorum formula ensures tolerance of f faulty agents.

4. Async Scalability: Ava’s concurrency model (TaskGroup + Semaphore) validates the async approach already used in Hermes infrastructure.

Actions

• 🟢 P2: Review Ava’s tiered memory implementation for gbrain lessons
• 🟡 P1: Consider emergence detection for cross-agent behavior monitoring
• 🟢 P2: Monitor Ava’s trajectory (400⭐ in first day → strong community interest)